A context-aware temporal modeling through unified multi-scale temporal encoding and hierarchical sequence learning for single-channel EEG sleep staging.

A single-channel EEG model improved automatic sleep-stage classification on public benchmark recordings, including the difficult N1 stage. The results support simpler monitoring research but are not a substitute for clinical polysomnography or prospective validation.

Vakili AA, Jahanshiri S, Salimi-Badr A.
PLOS ONE · Publication date: 2026-09-30
Added to site: 2026-10-01
DOI: 10.1371/journal.pone.0358241

Electronic publication date verified in PubMed, Crossref and the full text. This source date is separate from the site's addition date.

Modalities

EEG, AI, Wearable EEG

Clinical focus

Sleep staging, Single-channel EEG, Temporal modeling

Population

Healthy participants represented in two public Sleep-EDF benchmark datasets

Sample

SleepEDF-20: 20 participants and 39 recordings; SleepEDF-78: 78 participants and 197 recordings

Intervention / method

TempoSleep used one Fpz-Cz EEG channel, multiscale convolution, temporal compression, hierarchical BiLSTM modeling, attention pooling, augmentation and class-weighted loss

Comparator

Published single-channel sleep-staging baselines and ablated versions of the proposed model

Clinician summary

Under five-fold grouped cross-validation on SleepEDF-20, TempoSleep reported 89.72% accuracy, 85.46% macro-F1 and N1 F1 of 61.7%. The full text also evaluated the larger SleepEDF-78 benchmark and used subject grouping to reduce epoch leakage.

How to interpret the findings

This was retrospective model development on two related public benchmark datasets rather than external prospective clinical validation. Only the Fpz-Cz channel was used, annotations carry expert uncertainty, and performance on clinical populations, contemporary wearable hardware and edge deployment remains unknown. Multimodal polysomnography remains the clinical reference standard.

Equipment and disclosures

The analysis used the Fpz-Cz channel from public Sleep-EDF polysomnography recordings sampled at 100 Hz. It did not validate a commercial wearable or diagnostic device.

The authors declared no competing interests and received no specific funding. The Sleep-EDF data are publicly available through PhysioNet.

Related evidence

Compare findings from the same clinical area, including mixed results.